Yuhan Helena Liu
Princeton University
hyliu24@uw.edu
Bio
I completed my PhD in Applied Mathematics at the University of Washington in June 2024 and am now a postdoc at Princeton University, working at the intersection of computational neuroscience and deep learning. My research leverages deep learning theory and large-scale neural recordings to explore how the brain processes information, funded by fellowships such as NSERC PGS-D, FRQNT B2X, Pearson, and NSF AccelNet IN-BIC. I have published as lead author in NeurIPS, ICLR, PNAS, and IEEE, focusing on biologically plausible learning models and their generalization properties. In addition to my theory-driven work, I use data-driven approaches to uncover learning strategies from animal data. I’ve also held research roles at the Allen Institute, Mila, and MIT, and have taught and mentored extensively, earning a departmental teaching award and being named a 2024 Rising Star in Computational and Data Sciences.
Areas of Research
- AI for Healthcare and Life Sciences
Deep learning frameworks for modeling how neural circuits learn
The brain’s prowess in learning and adapting remains an enigma, particularly in its approach to the ‘temporal credit assignment’ problem. How do neural circuits determine which specific states and connections contribute to future outcomes, and subsequently adjust these for enhanced learning? My research addresses this by combining insights from the latest large-scale neuroscience data and recent deep learning theoretical tools. I first introduce novel learning rules inspired by the Allen Institute’s transcriptomics data, which revealed widespread and intricate cell-type-specific interactions among neuromodulatory molecules. This rule enables neurons to propagate credit information efficiently, enhancing learning performance beyond that of biologically plausible predecessors. Extensive computational experiments confirm the significant role of local neuromodulatory signals in learning, offering new perspectives on neural information processing. I then assess the generalization capabilities of bio-plausible learning rules through the lens of deep learning theory, particularly focusing on the curvature of the loss landscape via the loss’ Hessian eigenspectrum. Our findings reveal that these rules often settle in high-curvature regions of the loss landscape, indicating suboptimal generalization. This analysis led to a mathematical theorem linking synaptic weight update dynamics to landscape curvature, proposing neuromodulator-driven adjustments as a potential enhancement for learning rule performance. Given how initial conditions can greatly influence a system’s future trajectory, I also delve into the impact of initial connectivity structures on learning dynamics in neural circuits. By examining various connectivity patterns derived from neuroscience data, our findings suggest that high-rank initializations leverage pre-existing high-dimensional input expansion to facilitate input decoding. This leads to minimal changes post-training, a higher propensity for lazy learning, and potential implications for increased metabolic costs and risks of catastrophic forgetting. These specific initializations thus predispose networks toward certain learning behaviors, critically affecting their ability to adapt and generalize.